CN112767279A - Underwater image enhancement method for generating countermeasure network based on discrete wavelet integration - Google Patents
Underwater image enhancement method for generating countermeasure network based on discrete wavelet integration Download PDFInfo
- Publication number
- CN112767279A CN112767279A CN202110134075.9A CN202110134075A CN112767279A CN 112767279 A CN112767279 A CN 112767279A CN 202110134075 A CN202110134075 A CN 202110134075A CN 112767279 A CN112767279 A CN 112767279A
- Authority
- CN
- China
- Prior art keywords
- image
- network
- convolution
- layer
- underwater
- Prior art date
- Legal status (The legal status is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the status listed.)
- Granted
Links
- 238000000034 method Methods 0.000 title claims abstract description 41
- 230000010354 integration Effects 0.000 title claims abstract description 14
- 238000013528 artificial neural network Methods 0.000 claims abstract description 46
- 238000010606 normalization Methods 0.000 claims abstract description 25
- 238000012549 training Methods 0.000 claims abstract description 17
- 238000012545 processing Methods 0.000 claims abstract description 15
- 230000002708 enhancing effect Effects 0.000 claims abstract description 9
- 238000011176 pooling Methods 0.000 claims description 52
- 230000006870 function Effects 0.000 claims description 40
- 230000004913 activation Effects 0.000 claims description 21
- 238000000605 extraction Methods 0.000 claims description 12
- 238000005096 rolling process Methods 0.000 claims description 10
- 238000011478 gradient descent method Methods 0.000 claims description 6
- 238000005070 sampling Methods 0.000 claims description 6
- 238000004364 calculation method Methods 0.000 claims description 5
- 238000007781 pre-processing Methods 0.000 claims description 4
- 238000000354 decomposition reaction Methods 0.000 claims description 3
- 239000000306 component Substances 0.000 description 28
- 238000003384 imaging method Methods 0.000 description 9
- 238000006731 degradation reaction Methods 0.000 description 6
- 238000013135 deep learning Methods 0.000 description 5
- XLYOFNOQVPJJNP-UHFFFAOYSA-N water Substances O XLYOFNOQVPJJNP-UHFFFAOYSA-N 0.000 description 5
- 230000015556 catabolic process Effects 0.000 description 4
- 238000010586 diagram Methods 0.000 description 3
- 230000004438 eyesight Effects 0.000 description 3
- 230000009286 beneficial effect Effects 0.000 description 2
- 230000000694 effects Effects 0.000 description 2
- 230000003287 optical effect Effects 0.000 description 2
- 230000005540 biological transmission Effects 0.000 description 1
- 239000002131 composite material Substances 0.000 description 1
- 239000008358 core component Substances 0.000 description 1
- 238000001514 detection method Methods 0.000 description 1
- 238000011161 development Methods 0.000 description 1
- 238000005516 engineering process Methods 0.000 description 1
- 239000012535 impurity Substances 0.000 description 1
- 238000005259 measurement Methods 0.000 description 1
- 230000003595 spectral effect Effects 0.000 description 1
- 230000000007 visual effect Effects 0.000 description 1
- 230000016776 visual perception Effects 0.000 description 1
Images
Classifications
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06T—IMAGE DATA PROCESSING OR GENERATION, IN GENERAL
- G06T5/00—Image enhancement or restoration
- G06T5/73—Deblurring; Sharpening
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06T—IMAGE DATA PROCESSING OR GENERATION, IN GENERAL
- G06T5/00—Image enhancement or restoration
- G06T5/10—Image enhancement or restoration using non-spatial domain filtering
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06T—IMAGE DATA PROCESSING OR GENERATION, IN GENERAL
- G06T5/00—Image enhancement or restoration
- G06T5/90—Dynamic range modification of images or parts thereof
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06T—IMAGE DATA PROCESSING OR GENERATION, IN GENERAL
- G06T2207/00—Indexing scheme for image analysis or image enhancement
- G06T2207/20—Special algorithmic details
- G06T2207/20048—Transform domain processing
- G06T2207/20064—Wavelet transform [DWT]
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06T—IMAGE DATA PROCESSING OR GENERATION, IN GENERAL
- G06T2207/00—Indexing scheme for image analysis or image enhancement
- G06T2207/20—Special algorithmic details
- G06T2207/20081—Training; Learning
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06T—IMAGE DATA PROCESSING OR GENERATION, IN GENERAL
- G06T2207/00—Indexing scheme for image analysis or image enhancement
- G06T2207/20—Special algorithmic details
- G06T2207/20084—Artificial neural networks [ANN]
Landscapes
- Physics & Mathematics (AREA)
- General Physics & Mathematics (AREA)
- Engineering & Computer Science (AREA)
- Theoretical Computer Science (AREA)
- Image Processing (AREA)
- Image Analysis (AREA)
Abstract
The invention relates to an underwater image enhancement method for generating a countermeasure network based on discrete wavelet integration. The method comprises the following steps: carrying out pairing processing, data enhancement and normalization processing on data to be trained, then zooming to a fixed size, and downsampling a label image, namely a real underwater image enhancement result to generate images with different sizes; designing a multi-level image enhancement generation network, and training an image enhancement model capable of enhancing the underwater image by using the designed network; designing a multi-level image identification neural network, and training an image identification model capable of predicting the probability that an input image is a real image by using the designed network; designing a generating network and identifying a neural network target loss function; alternately training the generating network and the discriminating neural network to converge to Nash balance by using the paired images; and inputting the underwater image into the trained image enhancement generation model, and outputting the enhanced image. The invention solves the problems of color distortion and blurring of the underwater image and can obviously improve the quality of the underwater image.
Description
Technical Field
The invention relates to the field of image processing and computer vision, in particular to an underwater image enhancement method for generating a countermeasure network based on discrete wavelet integration.
Background
The marine environment containing a large amount of resources is one of the core components of human sustainable development, however, the marine complex imaging environment causes the underwater image acquired by the optical vision system to be generally accompanied with various types of degradation, which brings great influence on underwater operation depending on vision, such as submarine resource exploration, marine archaeology, underwater target detection and the like. The main reason for the degradation of underwater images is the energy attenuation and scattering phenomena of light as it travels underwater. When light is transmitted underwater, energy attenuation is accompanied, the light attenuation degrees of different spectral regions are different, red light attenuation is the most serious, green light is the second, and blue light is the least. The underwater image generally appears to be bluish-green accompanied by a color deviation. The scattering of light in water is classified into forward scattering and backward scattering. Forward scattering refers to a scattering phenomenon that small angle deviation is generated when light reflected by an object in water is transmitted to a camera, and image details are blurred. Backscattering refers to the fact that when light rays irradiate on objects in water, the light rays encounter impurities in the water and are scattered and directly received by a camera, and the contrast of an image is reduced.
At present, some imaging devices specially designed for underwater imaging environments can overcome part of underwater imaging problems and acquire clear images. However, in practical application, the requirements for the quality of underwater images are far more than that, the requirements are still difficult to meet only by means of acquisition equipment, and advanced underwater imaging equipment has extremely high cost, so that the underwater imaging equipment is difficult to be widely applied to various underwater tasks. Compared with expensive underwater imaging equipment, the image processing technology has the advantages of low cost, easiness in operation and the like.
The existing underwater image enhancement methods are mainly divided into two types: firstly, mathematical modeling is carried out on the degradation process of an underwater image, the degradation process is inverted to obtain a clear underwater image by estimating model parameters, the method needs to establish a model according to information such as optical parameters of a water body, camera parameters and the distance between a camera and a target object, the information needs to be measured by manual measurement or other methods to obtain numerical values, the obtained information is generally inaccurate, and therefore the enhanced image is prone to the problems of detail blurring, color deviation and the like. Meanwhile, although the methods can achieve good effects in certain fixed scenes, underwater scenes are complex and changeable, underwater image degradation types are various, the methods are difficult to match all the degradation types, generalization capability is poor, and when the scenes are changed greatly, the models need to be built again.
Secondly, a method based on deep learning is used, the method is a data-driven method, has strong learning characteristic representation capability, is lower in complexity than a non-deep learning method, and is widely applied to the field of image processing, so that a plurality of underwater image enhancement algorithms based on deep learning are proposed in recent years. However, although the existing underwater image enhancement method based on deep learning can effectively improve the quality of underwater images, the existing underwater image enhancement method based on deep learning has shortcomings in local and detail enhancement of underwater images due to the fact that the underwater environment is too complex, and the recovered images have problems of color deviation, detail blurring and the like.
In summary, the underwater image is complex in imaging environment and low in imaging image quality, and the existing method is easy to generate the phenomena of detail blurring and color deviation on the result of underwater image enhancement. The method comprises the steps of dividing an image into high-frequency information and low-frequency information and processing the high-frequency information and the low-frequency information separately, effectively reducing information loss by utilizing the advantages of discrete wavelets in the aspect of signal reconstruction, generating an underwater image enhancement result which is more in line with human subjective visual effect through countertraining of a generation model and an identification model, and remarkably improving the quality of the underwater image.
Disclosure of Invention
The invention aims to provide an underwater image enhancement method for generating a countermeasure network based on discrete wavelet integration, which is beneficial to improving the quality of underwater images.
In order to achieve the purpose, the technical scheme of the invention is as follows: an underwater image enhancement method based on discrete wavelet integration generation countermeasure network comprises the following steps:
step S1, data preprocessing is carried out on the data to be trained: firstly, matching an underwater image with a corresponding label image, then performing data enhancement and normalization processing, then scaling the normalized image to a fixed size, and downsampling the label image, namely a real underwater image enhancement result, to generate images with different sizes;
s2, designing a multi-level image enhancement generation network, and training an image enhancement model for enhancing the underwater image by using the designed network;
step S3, designing a multi-level image identification neural network, and using the designed network to train an image identification model for predicting the probability that an input image is a real image;
step S4, designing a generating network and identifying a neural network target loss function;
step S5, alternately training the generation network and the discrimination neural network by using the paired images to converge to Nash balance;
and step S6, inputting the underwater image into the trained image enhancement generation model, and outputting the enhanced image.
In an embodiment of the present invention, the step S1 is implemented as follows:
step S11, matching the underwater image with the corresponding label image;
step S12, carrying out uniform random turning operation on all paired images to be trained, and enhancing data;
step S13, all the enhanced images to be trained are normalized, an image I (I, j) is given, and a normalization value is calculated
The formula of (1) is as follows:
wherein (i, j) represents the position of the pixel;
step S14, scaling all normalized images to a fixed size H multiplied by W;
step S15, down-sampling the label image to generate different size images, and giving image G(H,W)A label image G of H × W size is represented, and an image G 'is generated by down-sampling'kK is 2,3, wherein
Downsampling uses the nearest neighbor interpolation algorithm, and G'1G; for any label image G, label image sets { G 'of different sizes are obtained'1，G′2，G′3}。
In an embodiment of the present invention, the step S2 is implemented as follows:
step S21, designing a multi-level neural network structure for extracting effective features of images, inputting the feature extraction network into an underwater image, wherein the network comprises a plurality of convolution blocks and wavelet pooling layers; the convolution blocks consist of convolution layers, a normalization layer and an activation layer, the feature extraction network has three layers, each layer comprises a plurality of convolution blocks and a wavelet pooling layer, and the number of the convolution blocks in each layer can be different; the convolution layer uses convolution with convolution kernel of 3x3 and step length of 1, the normalization layer uses batch normalization, and the activation layer uses ReLu function; the wavelet pooling layer uses discrete Haar wavelets to decompose the features, using four decomposition kernels LLT,LHT,HLT,HHTWherein the low frequency filter and the high frequency filter are respectively
The wavelet pooling layer decomposes the input features into low and high frequency components, llk,lhk,hlk,hhkK is 1,2,3, wherein llkFor low frequency components, lhk,hlk,hhkAre all high frequency components; the low-frequency components are transmitted to the next level of the network, and each high-frequency component is directly transmitted into the image reconstruction network after passing through a feature thinning processing module; the feature refinement processing module consists of a plurality of residual attention modules, each residual attention module comprises two layers of convolution kernels of 3x3 and step lengthConvolution of 1 and a channel attention module; the output of the final feature extraction network is ll3Features and three levels of high frequency components lhk,hlk,hhk,k＝1,2,3；
Step S22, designing a multi-level neural network structure for image reconstruction, wherein the input of the image reconstruction network is ll extracted in the step S213Features and three levels of high frequency components lhk,hlk,hhkH is 1,2, 3; the image reconstruction network comprises a plurality of volume blocks, reconstruction blocks and a wavelet anti-pooling layer; the convolution block consists of a convolution layer, a normalization layer and an activation layer, and the reconstruction block consists of a convolution layer and an activation layer; the image reconstruction network has three levels which respectively correspond to the three levels of the feature extraction network, each level comprises a rolling block, a wavelet anti-pooling layer and a reconstruction block, and the number of the rolling blocks in each level can be different; convolution layers in the convolution blocks use convolution with convolution kernels of 3x3 and step length of 1, the normalization layer uses batch normalization, and the activation layer uses a ReLu function; convolution layers in the reconstruction blocks use convolution with convolution kernels of 3x3 and step length of 1, and activation layers use Tanh functions; the first level of input is ll3Features, low-frequency component ll is obtained after convolution block4Features as low frequency components of the wavelet anti-pooling layer; discrete Haar wavelet pair ll is used by wavelet anti-pooling layer4Characteristic and high frequency components lhk,hlk,hhkThe k is 3, and the output of the wavelet anti-pooling layer is used as the input of the reconstruction block on one hand and as the input characteristic of the second layer on the other hand; the second layer is rolled to obtain a low-frequency component ll5Characteristic, wavelet anti-pooling layer pair ll5Characteristic high frequency component lhk,hlk,hhkK is 2; similarly to the previous layer, the output of the wavelet anti-pooling layer is used as the input of the reconstruction block on one hand and as the input characteristic of the third layer on the other hand; the third layer is rolled to obtain a low-frequency component ll6Characteristic, wavelet anti-pooling layer pair ll6Characteristic high frequency component lhk,hlk,hhkCombining k with 1, and generating an H multiplied by W underwater image enhanced image through a rolling block and a reconstruction block; in addition, the method can be used for producing a composite materialThe features after the first two layers are combined generate an underwater image enhancement image with the current size through a reconstruction block; final output of three sizes of enhanced image I'kK is 1,2 and 3, and the sizes are H multiplied by W,1/2H multiplied by 1/2W,1/4H multiplied by 1/4W and G 'respectively'kAnd k is equal to 1,2 and 3.
In an embodiment of the present invention, the step S3 is specifically implemented as follows:
designing a multi-level neural network structure for image identification, wherein the image identification network inputs three-size enhanced images I 'generated by a generation network'kK is 1,2,3 or tag image G'kWhen k is 1,2,3, the network needs to be the same type of image for each input, namely, the same enhanced image or the same label image; the image identification network comprises a plurality of rolling blocks, a splicing layer and a pooling layer; the convolution block consists of a convolution layer and an activation layer, wherein the convolution layer uses convolution with a convolution kernel of 3x3 and a step length of 1, the activation layer uses a LeakyReLu function, and the pooling layer uses convolution with a convolution kernel of 4x4 and a step length of 2; image authentication network input l'1Or G'1Representing the underwater image and the label image respectively, firstly obtaining a characteristic vector d with the size of 1/2H multiplied by 1/2W through a pooling layer and two convolution blocks2And then picture I 'is processed using one convolution block'2Or G'2Extracting corresponding dimension characteristics f2The two characteristics are fused by using a characteristic splicing layer in a way of
Wherein
Representing a splicing operation; the fused features are subjected to a pooling layer and two convolution blocks to obtain a feature vector d with the size of 1/4H multiplied by 1/4W3And then picture I 'is processed using one convolution block'3Or G'3Extracting corresponding dimension characteristics f3Fusing the two characteristics by using a characteristic splicing layer in a fusing modeIs consistent with the previous layer; the fused feature vector passes through a pooling layer and two convolution blocks to obtain a feature vector with the size of 1/8H multiplied by 1/8W, and finally passes through a convolution layer with 1 channel to predict the probability of being a real image.
In an embodiment of the present invention, the step S4 is implemented as follows:
step S41, designing and generating a network target loss function, and generating a network total target loss function as follows:
wherein,
and
respectively L1 loss and generative countermeasure loss, λ1And λ2Is each loss balance coefficient, is a real number dot product operation; the specific calculation formula of each loss is as follows:
the generation network outputs three sizes of enhanced images, and the L1 loss only calculates the loss between the maximum size generated image and the reference label image; i'1Is to generate the largest size image, G 'in the network enhancement result'1Is the maximum size image in the reference label image, | | |1Is an absolute value operation;
wherein z represents an underwater image, Pz(z) probability distribution of underwater image, G generation network, D identification network, G (z) enhancement result of underwater image by generation network, D (G:)z)) represents the identification result of the output of the generation network by the identification network, a constant c is a numerical value used by the identification network for judging the difference between the generated image and the label image, wherein the value is c is 1, the generation network outputs enhanced images with three sizes in total, and all the enhanced images need to be sent to the identification network for calculating loss;
step S42, designing and identifying a neural network target loss function, wherein the identifying network target loss function is as follows:
where x denotes a reference label image, Pdata(x) Representing the probability distribution of the reference label image, z representing the underwater image, Pz(z) represents the probability distribution of the underwater image, G represents a generation network, D represents an identification network, D (x) represents the identification result of the identification neural network to the label image, G (z) represents the enhancement result of the generation network to the underwater image, D (G (z)) represents the identification result of the identification neural network to the output of the generation network, constants a and b represent the label image and the mark of the generation image respectively, the value is 1, b is 0, the generation network outputs the enhanced images of three sizes in total, all the enhanced images of three sizes corresponding to the reference label image need to be sent to the identification neural network to calculate loss, and all the enhanced images of three sizes corresponding to the reference label image need to be sent to the identification neural network to calculate loss.
In an embodiment of the present invention, the step S5 is implemented as follows:
step S51, randomly dividing the matched underwater images and label images into a plurality of batches, wherein each batch comprises N pairs of images;
step S52, inputting the underwater image into a generation network to obtain enhanced images I 'of three sizes'kK is 1,2,3, and the size is H × W,1/2H × 1/2W,1/4H × 1/4W in sequence, wherein H × W is the size specified in step S1;
step S53, inputting the generated enhanced image and the label image into an identifying neural network to obtain a judgment result;
step S54, calculating the gradient of each parameter in the generated network by using a back propagation method according to the total target loss function of the generated network, and updating the parameters of the generated network by using a random gradient descent method;
step S55, calculating the gradient of each parameter in the identified neural network by using a back propagation method according to the target loss function of the identified neural network, and updating the parameter of the identified neural network by using a random gradient descent method;
and S56, repeating the steps S51 to S54 of generating the network and identifying the neural network training by taking batches as units until the target loss function value of the generated network and the target loss function value of the identifying neural network converge to Nash balance, storing the network parameters, and finishing the training process of the generated network and the identifying neural network.
Compared with the prior art, the invention has the following beneficial effects:
the method is suitable for underwater image enhancement in various complex environments, can effectively restore the distorted color of the image, remove the image blur, and improve the contrast and brightness of the image. The enhanced image conforms to human subjective visual perception. The invention provides an underwater image enhancement network based on discrete wavelets, which can effectively reduce information loss in the image transmission process, retain image detail information, avoid detail blurring and be suitable for most complex scenes.
Drawings
FIG. 1 is a flow chart of an implementation of the method of the present invention.
Fig. 2 is a diagram of a network model architecture in an embodiment of the present invention.
Fig. 3 is a structural diagram of a feature refinement module in an embodiment of the present invention.
FIG. 4 is a diagram of a residual attention module in the feature refinement module according to an embodiment of the present invention.
Detailed Description
The technical scheme of the invention is specifically explained below with reference to the accompanying drawings.
The invention relates to an underwater image enhancement method for generating a countermeasure network based on discrete wavelet integration, which comprises the following steps:
step S1, data preprocessing is carried out on the data to be trained: firstly, matching an underwater image with a corresponding label image, then performing data enhancement and normalization processing, then scaling the normalized image to a fixed size, and downsampling the label image, namely a real underwater image enhancement result, to generate images with different sizes;
s2, designing a multi-level image enhancement generation network, and training an image enhancement model for enhancing the underwater image by using the designed network;
step S3, designing a multi-level image identification neural network, and using the designed network to train an image identification model for predicting the probability that an input image is a real image;
step S4, designing a generating network and identifying a neural network target loss function;
step S5, alternately training the generation network and the discrimination neural network by using the paired images to converge to Nash balance;
and step S6, inputting the underwater image into the trained image enhancement generation model, and outputting the enhanced image.
The following is a specific implementation of the present invention.
As shown in fig. 1, an underwater image enhancement method for generating a countermeasure network based on discrete wavelet integration includes the following steps:
s1, preprocessing data to be trained, firstly pairing the data, then enhancing and normalizing the data, then scaling the normalized image to a fixed size, and downsampling a label image, namely a real underwater image enhancement result to generate images with different sizes;
s2, designing a multi-level image enhancement generation network, and training an image enhancement model capable of enhancing the underwater image by using the designed network;
step S3, designing a multi-level image identification neural network, and using the designed network to train an image identification model capable of predicting the probability that an input image is a real image;
step S4, designing and generating a network and identifying a network target loss function;
step S5, alternately training the generation network and the identification network by using the paired images to converge to Nash balance;
and step S6, inputting the underwater image into the trained image enhancement generation model, and outputting the enhanced image.
Further, step S1 includes the steps of:
step S11: and matching the underwater image with the corresponding label image.
Step S12: and carrying out uniform random overturning operation on all paired images to be trained, and enhancing the data.
Step S13: all images to be trained are normalized, an image I (I, j) is given, and a normalization value is calculated
The formula of (1) is as follows:
where (i, j) represents the position of the pixel.
Step S14, scale all normalized images to a fixed size H × W.
Step S15, down-sampling the label image to generate different size images, and giving image G(H,W)A label image G of H × W size is represented, and an image G 'is generated by down-sampling'kK is 2,3, wherein
The downsampling uses a nearest neighbor interpolation algorithm. And G'1G. For any label image G, label image sets { G 'of different sizes are obtained'1，G′2，G′3}。
Further, step S2 includes the steps of:
s21, designing a multi-level neural network structure for extracting effective features of the image, wherein the feature extraction network is input as an underwater imageLike this, the network includes a plurality of volume blocks and a wavelet pooling layer. The convolution blocks are composed of convolution layers, a normalization layer and an activation layer, the feature extraction network has three layers, each layer comprises a plurality of convolution blocks and a wavelet pooling layer, and the number of the convolution blocks in each layer can be different. The convolutional layer uses convolution with a convolution kernel of 3x3 and a step size of 1, the normalization layer uses batch normalization, and the activation layer uses the ReLu function. The wavelet pooling layer uses discrete Haar wavelets to decompose the features, using four decomposition kernels LLT,LHT,HLT,HHTWherein the high-frequency filters are respectively
The wavelet pooling layer decomposes the input features into low and high frequency components, llk,lhk,hlk,hhkK is 1,2,3, wherein llkFor low frequency components, lhk,hlk,hhkAre all high frequency components. The low-frequency components are transmitted to the next level of the network, and each high-frequency component is directly transmitted into the image reconstruction network after passing through a feature thinning processing module. The feature refinement processing module consists of a plurality of residual attention modules, each of which consists of two layers of convolution kernel of 3x3 with step size of 1 and one channel attention module. The output of the final feature extraction network is ll3Features and three levels of high frequency components lhk,hlk,hhk,k＝1,2,3。
Step S22, designing a multi-level neural network structure for image reconstruction, wherein the input of the image reconstruction network is ll extracted in the step S213Features and three levels of high frequency components lhk,hlk,hhkAnd k is 1,2, 3. The image reconstruction network includes a plurality of volume blocks, reconstruction blocks, and a wavelet anti-pooling layer. The convolution block is composed of a convolution layer, a normalization layer and an activation layer, and the reconstruction block is composed of a convolution layer and an activation layer. The image reconstruction network has three levels which respectively correspond to the three levels of the feature extraction network, each level comprises a rolling block, a wavelet anti-pooling layer and a reconstruction block, and the number of the rolling blocks in each level can be different. Convolutional layer in convolutional blockConvolution with a convolution kernel of 3x3 and a step size of 1 was used, the normalization layer used batch normalization, and the activation layer used the ReLu function. The convolutional layers in the reconstruction block use convolution with a convolution kernel of 3x3 and step size 1, and the active layers use the Tanh function. The first level of input is ll3Features that after several convolution blocks, the low-frequency component ll is obtained4And the characteristic is used as the low-frequency component of the wavelet anti-pooling layer. Discrete Haar wavelet pair ll is used by wavelet anti-pooling layer4Characteristic and high frequency components lhk,hlk,hhkAnd k is 3, the inverse pooling kernel parameters are the same as those in step S21, and the output of the wavelet inverse pooling layer is used as the input of the reconstruction block on one hand and as the input feature of the second layer on the other hand. The second layer is rolled to obtain a low-frequency component ll5Characteristic, wavelet anti-pooling layer pair ll5Characteristic high frequency component lhk,hlk,hhkAnd k is 2. Similarly to the previous layer, the output of the wavelet anti-pooling layer is used as the input of the reconstruction block on one hand and as the input feature of the third layer on the other hand. The third layer is rolled to obtain a low-frequency component ll6Characteristic, wavelet anti-pooling layer pair ll6Characteristic high frequency component lhk,hlk,hhkAnd k is combined to be 1, and then the H multiplied by W underwater image enhanced image is generated through several volume blocks and a reconstruction block. In addition, the features of the first two layers after combination are subjected to reconstruction to generate an underwater image enhanced image with the current size. Final output of three sizes of enhanced image I'kK is 1,2 and 3, and the sizes are H multiplied by W,1/2H multiplied by 1/2W,1/4H multiplied by 1/4W and G 'respectively'kAnd k is equal to 1,2, and 3, wherein H × W is the size specified in step S1.
Further, step S3 is implemented as follows:
designing a multi-level neural network structure for image identification, wherein the image identification network inputs three-size enhanced images I 'generated by a generation network'kK is 1,2,3 or tag image G'kAnd k is 1,2 and 3, the network needs to be the same type of image, namely the enhanced image or the label image, every time the network inputs the image, and the image size is the size specified in the step S1 and the step S2. Image authentication networkComprising a plurality of volume blocks, a splice layer and a pooling layer. The convolutional block consists of convolutional layers and active layers, wherein the convolutional layers use convolution with a convolution kernel of 3x3 and a step length of 1, the active layers use LeakyReLu functions, and the pooling layers use convolution with a convolution kernel of 4x4 and a step length of 2. Network inputs I'1Or G'1Representing the underwater image and the label image respectively, firstly obtaining a characteristic vector d with the size of 1/2H multiplied by 1/2W through a pooling layer and two convolution blocks2And then picture I 'is processed using one convolution block'2Or G'2Extracting corresponding dimension characteristics f2The two characteristics are fused by using a characteristic splicing layer in a way of
Wherein
Indicating a splicing operation. The fused features are subjected to a pooling layer and two convolution blocks to obtain a feature vector d with the size of 1/4H multiplied by 1/4W3And then picture I 'is processed using one convolution block'3Or G'3Extracting corresponding dimension characteristics f3And fusing the two characteristics by using the characteristic splicing layer, wherein the fusing mode is consistent with that of the previous layer. The fused feature vector passes through a pooling layer and two convolution blocks to obtain a feature vector with the size of 1/8H multiplied by 1/8W, and finally passes through a convolution layer with 1 channel to predict the probability of being a real image.
Further, step S4 includes the steps of:
step S41, designing and generating a network target loss function, and generating a network total target loss function as follows:
wherein,
and
respectively L1 loss and generative countermeasure loss, λ1And λ2Is each loss balance coefficient, is a real number dot product operation; the specific calculation formula of each loss is as follows:
the generation network outputs three sizes of enhanced images, and the L1 penalty only calculates the penalty between the largest size generated image and the reference label image. I'1Is to generate the largest size image, G 'in the network enhancement result'1Is the maximum size image in the reference label image, | | |1Is an absolute value operation;
wherein z represents an underwater image, Pz(z) represents the probability distribution of the underwater image, G represents a generation network, D represents an identification network, G (z) represents the enhancement result of the generation network on the underwater image, D (G (z)) represents the output identification result of the identification network on the generation network, a constant c is a numerical value used by the identification network for judging the difference between the generation image and the label image, the value is c-1, the generation network outputs the enhancement images with three sizes in total, and all the enhancement images need to be sent to the identification network for calculating loss.
Step S42, designing and identifying a network target loss function, wherein the network target loss function is as follows:
where x denotes a reference label picture, Pdata(x) Representing the probability distribution of the reference label image, z representing the underwater image, Pz(z)
The method comprises the steps of representing underwater image probability distribution, G representing a generation network, D representing an identification network, D (x) representing identification results of the identification network on label images, G (z) representing enhancement results of the generation network on the underwater images, D (G (z)) representing identification results of the identification network on output of the generation network, constants a and b respectively representing label images and marks of the generation images, wherein the values are a-1, b-0, the generation network outputs three-size enhancement images in total, all the three-size enhancement images need to be sent to the identification network for calculation loss, and all the three-size images corresponding to reference label images need to be sent to the identification network for calculation loss.
Further, step S5 includes the steps of:
step S51, randomly dividing the matched underwater images and label images into a plurality of batches, wherein each batch comprises N pairs of images;
step S52, inputting the underwater image into the generation network of step S2 to obtain three-size enhanced images I'kAnd k is 1,2 and 3, and the sizes are H multiplied by W,1/2H multiplied by 1/2W and 1/4H multiplied by 1/4W in sequence, wherein H multiplied by W is the size specified in the step S1.
Step S53, the generated enhanced image and label image are input to the discrimination network in step S3 to obtain the discrimination result.
Step S54, calculating the gradient of each parameter in the generated network by using a back propagation method according to the total target loss function of the generated network, and updating the parameters of the generated network by using a random gradient descent method;
step S55, calculating the gradient of each parameter in the identification network by using a back propagation method according to the target loss function of the identification network, and updating the parameters of the identification network by using a random gradient descent method;
and S56, repeating the steps from S51 to S54 for generating the network and identifying the network training by taking batches as units until the target loss function value of the generated network and the target loss function value of the identified network converge to Nash balance, storing the network parameters, and finishing the training process of the generated network and the identified network.
The above are preferred embodiments of the present invention, and all changes made according to the technical scheme of the present invention that produce functional effects do not exceed the scope of the technical scheme of the present invention belong to the protection scope of the present invention.
Claims (6)
1. An underwater image enhancement method for generating a countermeasure network based on discrete wavelet integration is characterized by comprising the following steps:
step S1, data preprocessing is carried out on the data to be trained: firstly, matching an underwater image with a corresponding label image, then performing data enhancement and normalization processing, then scaling the normalized image to a fixed size, and downsampling the label image, namely a real underwater image enhancement result, to generate images with different sizes;
s2, designing a multi-level image enhancement generation network, and training an image enhancement model for enhancing the underwater image by using the designed network;
step S3, designing a multi-level image identification neural network, and using the designed network to train an image identification model for predicting the probability that an input image is a real image;
step S4, designing a generating network and identifying a neural network target loss function;
step S5, alternately training the generation network and the discrimination neural network by using the paired images to converge to Nash balance;
and step S6, inputting the underwater image into the trained image enhancement generation model, and outputting the enhanced image.
2. The underwater image enhancement method for generating the countermeasure network based on the discrete wavelet integration according to claim 1, wherein the step S1 is implemented by the following steps:
step S11, matching the underwater image with the corresponding label image;
step S12, carrying out uniform random turning operation on all paired images to be trained, and enhancing data;
step S13, all the enhanced images to be trained are normalized, an image I (I, j) is given, and a normalization value is calculated
Is disclosedThe formula is as follows:
wherein (i, j) represents the position of the pixel;
step S14, scaling all normalized images to a fixed size H multiplied by W;
step S15, down-sampling the label image to generate different size images, and giving image G(H,W)A label image G of H × W size is represented, and an image G 'is generated by down-sampling'kK is 2,3, wherein
Downsampling uses the nearest neighbor interpolation algorithm, and G'1G; for any label image G, label image sets { G 'of different sizes are obtained'1，G′2，G′3}。
3. The underwater image enhancement method for generating the countermeasure network based on the discrete wavelet integration according to claim 2, wherein the step S2 is implemented by the following steps:
step S21, designing a multi-level neural network structure for extracting effective features of images, inputting the feature extraction network into an underwater image, wherein the network comprises a plurality of convolution blocks and wavelet pooling layers; the convolution blocks consist of convolution layers, a normalization layer and an activation layer, the feature extraction network has three layers, each layer comprises a plurality of convolution blocks and a wavelet pooling layer, and the number of the convolution blocks in each layer can be different; the convolution layer uses convolution with convolution kernel of 3x3 and step length of 1, the normalization layer uses batch normalization, and the activation layer uses ReLu function; the wavelet pooling layer uses discrete Haar wavelets to decompose the features, using four decomposition kernels LLT,LHT,HLT,HHTWherein the low frequency filter and the high frequency filter are respectively
The wavelet pooling layer decomposes the input features into low and high frequency components, llk,lhk,hlk,hhkK is 1,2,3, wherein llkFor low frequency components, lhk,hlk,hhkAre all high frequency components; the low-frequency components are transmitted to the next level of the network, and each high-frequency component is directly transmitted into the image reconstruction network after passing through a feature thinning processing module; the characteristic thinning processing module consists of a plurality of residual error attention modules, and each residual error attention module consists of two layers of convolution kernels with the convolution rate of 3x3 and the step length of 1 and a channel attention module; the output of the final feature extraction network is ll3Features and three levels of high frequency components lhk,hlk,hhk,k＝1,2,3；
Step S22, designing a multi-level neural network structure for image reconstruction, wherein the input of the image reconstruction network is ll extracted in the step S213Features and three levels of high frequency components lhk,hlk,hhkK is 1,2, 3; the image reconstruction network comprises a plurality of volume blocks, reconstruction blocks and a wavelet anti-pooling layer; the convolution block consists of a convolution layer, a normalization layer and an activation layer, and the reconstruction block consists of a convolution layer and an activation layer; the image reconstruction network has three levels which respectively correspond to the three levels of the feature extraction network, each level comprises a rolling block, a wavelet anti-pooling layer and a reconstruction block, and the number of the rolling blocks in each level can be different; convolution layers in the convolution blocks use convolution with convolution kernels of 3x3 and step length of 1, the normalization layer uses batch normalization, and the activation layer uses a ReLu function; convolution layers in the reconstruction blocks use convolution with convolution kernels of 3x3 and step length of 1, and activation layers use Tanh functions; the first level of input is ll3Features, low-frequency component ll is obtained after convolution block4Features as low frequency components of the wavelet anti-pooling layer; discrete Haar wavelet pair ll is used by wavelet anti-pooling layer4Characteristic and high frequency components lhk,hlk,hhkThe k is 3, and the output of the wavelet anti-pooling layer is used as the input of the reconstruction block on one hand and as the input characteristic of the second layer on the other hand; the second layer is rolled to obtain a low-frequency component ll5Characteristic, wavelet anti-pooling layer pair ll5Characteristic high frequency component lhk,hlk,hhkK is 2; similarly to the previous layer, the output of the wavelet anti-pooling layer is used as the input of the reconstruction block on one hand and as the input characteristic of the third layer on the other hand; the third layer is rolled to obtain a low-frequency component ll6Characteristic, wavelet anti-pooling layer pair ll6Characteristic high frequency component lhk,hlk,hhkCombining k with 1, and generating an H multiplied by W underwater image enhanced image through a rolling block and a reconstruction block; in addition, the features of the first two layers after combination generate an underwater image enhanced image with the current size through a reconstruction block; final output of three sizes of enhanced image I'kK is 1,2 and 3, and the sizes are H multiplied by W,1/2H multiplied by 1/2W,1/4H multiplied by 1/4W and G 'respectively'kAnd k is equal to 1,2 and 3.
4. The underwater image enhancement method for generating a countermeasure network based on discrete wavelet integration according to claim 3, wherein the step S3 is implemented as follows:
designing a multi-level neural network structure for image identification, wherein the image identification network inputs three-size enhanced images I 'generated by a generation network'kK is 1,2,3 or tag image G'kWhen k is 1,2,3, the network needs to be the same type of image for each input, namely, the same enhanced image or the same label image; the image identification network comprises a plurality of rolling blocks, a splicing layer and a pooling layer; the convolution block consists of a convolution layer and an activation layer, wherein the convolution layer uses convolution with a convolution kernel of 3x3 and a step length of 1, the activation layer uses a LeakyReLu function, and the pooling layer uses convolution with a convolution kernel of 4x4 and a step length of 2; image authentication network input l'1Or G'1Representing the underwater image and the label image respectively, firstly obtaining a characteristic vector d with the size of 1/2H multiplied by 1/2W through a pooling layer and two convolution blocks2And then picture I 'is processed using one convolution block'2Or G'2Extracting corresponding dimension characteristics f2The two characteristics are fused by using a characteristic splicing layer in a way of
Wherein
Representing a splicing operation; the fused features are subjected to a pooling layer and two convolution blocks to obtain a feature vector d with the size of 1/4H multiplied by 1/4W3And then picture I 'is processed using one convolution block'3Or G'3Extracting corresponding dimension characteristics f3Fusing the two characteristics by using a characteristic splicing layer, wherein the fusing mode is consistent with that of the previous layer; the fused feature vector passes through a pooling layer and two convolution blocks to obtain a feature vector with the size of 1/8H multiplied by 1/8W, and finally passes through a convolution layer with 1 channel to predict the probability of being a real image.
5. The underwater image enhancement method for generating the countermeasure network based on discrete wavelet integration according to claim 4, wherein the step S4 is implemented by the following steps:
step S41, designing and generating a network target loss function, and generating a network total target loss function as follows:
wherein,
and
respectively L1 loss and generative countermeasure loss, λ1And λ2Is each loss balance coefficient, is a real number dot product operation; the specific calculation formula of each loss is as follows:
the generation network outputs three sizes of enhanced images, and the L1 loss only calculates the loss between the maximum size generated image and the reference label image; i'1Is to generate the largest size image, G 'in the network enhancement result'1Is the maximum size image in the reference label image, | | |1Is an absolute value operation;
wherein z represents an underwater image, Pz(z) representing underwater image probability distribution, G representing a generation network, D representing an identification network, G (z) representing an enhancement result of the generation network on the underwater image, D (G (z)) representing an identification result of the identification network on the output of the generation network, and a constant c being a numerical value used by the identification network for judging the difference between the generation image and the label image, wherein the value is c being 1, the generation network outputs the enhancement images of three sizes in total, and the enhancement images are required to be sent to the identification network for calculating loss;
step S42, designing and identifying a neural network target loss function, wherein the identifying network target loss function is as follows:
where x denotes a reference label image, Pdata(x) Representing the probability distribution of the reference label image, z representing the underwater image, Pz(z) represents underwater image probability distribution, G represents generation network, D represents identification network, D (x) represents identification result of identification neural network to label image, G (z) represents enhancement result of generation network to underwater image, D (G (z)) represents identification result of identification neural network to output of generation network, constants a and b represent label image and label of generation image respectively, wherein, the values are a-1, b-0, and generation networkThe three-size enhanced images are output by the network, and all the three-size enhanced images need to be sent to the discrimination neural network for calculating loss, and all the three-size enhanced images corresponding to the reference label image need to be sent to the discrimination neural network for calculating loss.
6. The underwater image enhancement method for generating the countermeasure network based on discrete wavelet integration according to claim 5, wherein the step S5 is implemented by the following steps:
step S51, randomly dividing the matched underwater images and label images into a plurality of batches, wherein each batch comprises N pairs of images;
step S52, inputting the underwater image into a generation network to obtain enhanced images I 'of three sizes'kK is 1,2,3, and the size is H × W,1/2H × 1/2W,1/4H × 1/4W in sequence, wherein H × W is the size specified in step S1;
step S53, inputting the generated enhanced image and the label image into an identifying neural network to obtain a judgment result;
step S54, calculating the gradient of each parameter in the generated network by using a back propagation method according to the total target loss function of the generated network, and updating the parameters of the generated network by using a random gradient descent method;
step S55, calculating the gradient of each parameter in the identified neural network by using a back propagation method according to the target loss function of the identified neural network, and updating the parameter of the identified neural network by using a random gradient descent method;
and S56, repeating the steps S51 to S54 of generating the network and identifying the neural network training by taking batches as units until the target loss function value of the generated network and the target loss function value of the identifying neural network converge to Nash balance, storing the network parameters, and finishing the training process of the generated network and the identifying neural network.
Priority Applications (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN202110134075.9A CN112767279B (en) | 2021-02-01 | 2021-02-01 | Underwater image enhancement method for generating countermeasure network based on discrete wavelet integration |
Applications Claiming Priority (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN202110134075.9A CN112767279B (en) | 2021-02-01 | 2021-02-01 | Underwater image enhancement method for generating countermeasure network based on discrete wavelet integration |
Publications (2)
| Publication Number | Publication Date |
|---|---|
| CN112767279A true CN112767279A (en) | 2021-05-07 |
| CN112767279B CN112767279B (en) | 2022-06-14 |
Family
ID=75704416
Family Applications (1)
| Application Number | Title | Priority Date | Filing Date |
|---|---|---|---|
| CN202110134075.9A Active CN112767279B (en) | 2021-02-01 | 2021-02-01 | Underwater image enhancement method for generating countermeasure network based on discrete wavelet integration |
Country Status (1)
| Country | Link |
|---|---|
| CN (1) | CN112767279B (en) |
Cited By (7)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN113222953A (en) * | 2021-05-21 | 2021-08-06 | 西安电子科技大学 | Natural image enhancement method based on depth gamma transformation |
| CN113240606A (en) * | 2021-05-24 | 2021-08-10 | 博奥生物集团有限公司 | Traditional Chinese medicine inspection image color correction method and system |
| CN113284061A (en) * | 2021-05-17 | 2021-08-20 | 大连海事大学 | Underwater image enhancement method based on gradient network |
| CN114092793A (en) * | 2021-11-12 | 2022-02-25 | 杭州电子科技大学 | End-to-end biological target detection method suitable for complex underwater environment |
| CN114612330A (en) * | 2022-03-15 | 2022-06-10 | 中国人民解放军国防科技大学 | Image enhancement method of multiple cameras |
| CN114881879A (en) * | 2022-05-17 | 2022-08-09 | 燕山大学 | Underwater image enhancement method based on brightness compensation residual error network |
| CN117152445A (en) * | 2023-10-31 | 2023-12-01 | 暨南大学 | Real-time image semantic segmentation method and system based on multi-connection coding wavelet pooling |
Citations (7)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN109685072A (en) * | 2018-12-22 | 2019-04-26 | 北京工业大学 | A kind of compound degraded image high quality method for reconstructing based on generation confrontation network |
| US20190295223A1 (en) * | 2018-03-22 | 2019-09-26 | Adobe Inc. | Aesthetics-guided image enhancement |
| CN111242865A (en) * | 2020-01-10 | 2020-06-05 | 南京航空航天大学 | Fundus image enhancement method based on generation type countermeasure network |
| CN111275640A (en) * | 2020-01-17 | 2020-06-12 | 天津大学 | Image enhancement method for fusing two-dimensional discrete wavelet transform and generating countermeasure network |
| US20200286208A1 (en) * | 2019-03-08 | 2020-09-10 | International Business Machines Corporation | Neural network based enhancement of intensity images |
| CN111798400A (en) * | 2020-07-20 | 2020-10-20 | 福州大学 | Non-reference low-illumination image enhancement method and system based on generation countermeasure network |
| CN112288658A (en) * | 2020-11-23 | 2021-01-29 | 杭州师范大学 | Underwater image enhancement method based on multi-residual joint learning |
-
2021
- 2021-02-01 CN CN202110134075.9A patent/CN112767279B/en active Active
Patent Citations (7)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| US20190295223A1 (en) * | 2018-03-22 | 2019-09-26 | Adobe Inc. | Aesthetics-guided image enhancement |
| CN109685072A (en) * | 2018-12-22 | 2019-04-26 | 北京工业大学 | A kind of compound degraded image high quality method for reconstructing based on generation confrontation network |
| US20200286208A1 (en) * | 2019-03-08 | 2020-09-10 | International Business Machines Corporation | Neural network based enhancement of intensity images |
| CN111242865A (en) * | 2020-01-10 | 2020-06-05 | 南京航空航天大学 | Fundus image enhancement method based on generation type countermeasure network |
| CN111275640A (en) * | 2020-01-17 | 2020-06-12 | 天津大学 | Image enhancement method for fusing two-dimensional discrete wavelet transform and generating countermeasure network |
| CN111798400A (en) * | 2020-07-20 | 2020-10-20 | 福州大学 | Non-reference low-illumination image enhancement method and system based on generation countermeasure network |
| CN112288658A (en) * | 2020-11-23 | 2021-01-29 | 杭州师范大学 | Underwater image enhancement method based on multi-residual joint learning |
Non-Patent Citations (2)
| Title |
|---|
| QILI DENG ET AL.: "HardGAN: A Haze-Aware Representation Distillation GAN for Single Image Dehazing", 《COMPUTER VISION–ECCV 2020》 * |
| 沈亮: "复杂天气条件下水天场景图像增强技术研究", 《中国优秀博硕士学位论文全文数据库(硕士)工程科技Ⅱ辑》 * |
Cited By (13)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN113284061A (en) * | 2021-05-17 | 2021-08-20 | 大连海事大学 | Underwater image enhancement method based on gradient network |
| CN113284061B (en) * | 2021-05-17 | 2024-04-05 | 大连海事大学 | Underwater image enhancement method based on gradient network |
| CN113222953B (en) * | 2021-05-21 | 2023-10-20 | 西安电子科技大学 | Natural image enhancement method based on depth gamma transformation |
| CN113222953A (en) * | 2021-05-21 | 2021-08-06 | 西安电子科技大学 | Natural image enhancement method based on depth gamma transformation |
| CN113240606A (en) * | 2021-05-24 | 2021-08-10 | 博奥生物集团有限公司 | Traditional Chinese medicine inspection image color correction method and system |
| CN114092793A (en) * | 2021-11-12 | 2022-02-25 | 杭州电子科技大学 | End-to-end biological target detection method suitable for complex underwater environment |
| CN114092793B (en) * | 2021-11-12 | 2024-05-17 | 杭州电子科技大学 | End-to-end biological target detection method suitable for complex underwater environment |
| CN114612330B (en) * | 2022-03-15 | 2024-02-02 | 中国人民解放军国防科技大学 | Image enhancement method for multiple cameras |
| CN114612330A (en) * | 2022-03-15 | 2022-06-10 | 中国人民解放军国防科技大学 | Image enhancement method of multiple cameras |
| CN114881879A (en) * | 2022-05-17 | 2022-08-09 | 燕山大学 | Underwater image enhancement method based on brightness compensation residual error network |
| CN114881879B (en) * | 2022-05-17 | 2024-09-24 | 燕山大学 | Underwater image enhancement method based on brightness compensation residual error network |
| CN117152445A (en) * | 2023-10-31 | 2023-12-01 | 暨南大学 | Real-time image semantic segmentation method and system based on multi-connection coding wavelet pooling |
| CN117152445B (en) * | 2023-10-31 | 2024-01-12 | 暨南大学 | Real-time image semantic segmentation method and system based on multi-connection coding wavelet pooling |
Also Published As
| Publication number | Publication date |
|---|---|
| CN112767279B (en) | 2022-06-14 |
Similar Documents
| Publication | Publication Date | Title |
|---|---|---|
| CN112767279B (en) | Underwater image enhancement method for generating countermeasure network based on discrete wavelet integration | |
| CN112233038B (en) | True image denoising method based on multi-scale fusion and edge enhancement | |
| CN112507997B (en) | Face super-resolution system based on multi-scale convolution and receptive field feature fusion | |
| CN112767251B (en) | Image super-resolution method based on multi-scale detail feature fusion neural network | |
| CN108932536B (en) | Face posture reconstruction method based on deep neural network | |
| CN106548159A (en) | Reticulate pattern facial image recognition method and device based on full convolutional neural networks | |
| CN111861894A (en) | Image motion blur removing method based on generating type countermeasure network | |
| CN113284061B (en) | Underwater image enhancement method based on gradient network | |
| Liu et al. | Learning hadamard-product-propagation for image dehazing and beyond | |
| CN111179196B (en) | Multi-resolution depth network image highlight removing method based on divide-and-conquer | |
| Lu et al. | Underwater image enhancement method based on denoising diffusion probabilistic model | |
| CN116703752A (en) | Image defogging method and device of near infrared fused transducer structure | |
| CN112614070A (en) | DefogNet-based single image defogging method | |
| CN112164010A (en) | Multi-scale fusion convolution neural network image defogging method | |
| CN114004766A (en) | Underwater image enhancement method, system and equipment | |
| Pan et al. | ChebyLighter: Optimal Curve Estimation for Low-light Image Enhancement | |
| Zhuang et al. | A dense stereo matching method based on optimized direction-information images for the real underwater measurement environment | |
| CN115170944A (en) | NR underwater enhanced image quality evaluation method based on CNN | |
| CN113628143A (en) | Weighted fusion image defogging method and device based on multi-scale convolution | |
| CN114820395B (en) | Underwater image enhancement method based on multi-field information fusion | |
| CN107301625A (en) | Image defogging algorithm based on brightness UNE | |
| Huang et al. | Attention-based for multiscale fusion underwater image enhancement | |
| CN115272131B (en) | Image mole pattern removing system and method based on self-adaptive multispectral coding | |
| CN116703750A (en) | Image defogging method and system based on edge attention and multi-order differential loss | |
| Natarajan | A review on underwater image enhancement techniques |
Legal Events
| Date | Code | Title | Description |
|---|---|---|---|
| PB01 | Publication | ||
| PB01 | Publication | ||
| SE01 | Entry into force of request for substantive examination | ||
| SE01 | Entry into force of request for substantive examination | ||
| GR01 | Patent grant | ||
| GR01 | Patent grant |